v2.7.0

Adapt weight decay at deepspeed Adam optimizers.
Weight decay hparams tuning.
Model safetensors keep trained dtype.

v2.6.1

Requirements updated.

v2.6.0

Gather and convert as a model.pt after all deepspeed strategy training for uploading hub.
Postprocessing(uploading hub) modules and scripts.

v2.5.0

Fix: unused monitor parameter.

v2.4.3

Tuner refactoring.

v2.4.2

Delete strategy at get_architecture() in SetUp class.
Parameterize direction and monitor at tuner.
Update Packages.txt

v2.4.1

Adapt dot env at soft voting.

v2.4.0

T_max -> period(t_max = period * total steps per epoch), gradient clipping, dotenv, basic scripts.

v2.3.3

A detail of ckpt name changed (using "=") at main config file, specify huggingface cache path at main.py

v2.3.2

Add precision in save_detail, manage total_batch_size param.